Proceedings of the Tenth International AAAI Conference on
Web and Social Media (ICWSM 2016)
The Emotional and Chromatic Layers of Urban Smells
Daniele Quercia
Luca Maria Aiello
Rossano Schifanella
Bell Labs
quercia@acm.org
Yahoo Labs
alucca@yahoo-inc.com
University of Turin
schifane@di.unito.it
Flickr and Instagram tags). By building upon that methodology (Section 3), we recreate the smellscapes of London
and Barcelona. We capitalize on that mapping to make three
new contributions:
Abstract
People are able to detect up to 1 trillion odors. Yet,
city planning is concerned only with a few bad odors,
mainly because odors are currently captured only
through complaints made by urban dwellers. To capture
both good and bad odors, we resort to a methodology
that has been recently proposed and relies on tagging
information of geo-referenced pictures. In doing so for
the cities of London and Barcelona, this work makes
three new contributions. We study 1) how the urban
smellscape changes in time and space; 2) which emotions people share at places with specific smells; and 3)
what is the color of a smell, if it exists. Without social media data, insights about those three aspects have
been difficult to produce in the past, further delaying the
creation of urban restorative experiences.
1
• We study the temporal and spatial dynamics of smell
(Section 4). As for temporal aspects, we find that the most
seasonal smells are those of plant and nature, and the most
olfactory pleasant months are in Spring. As for spatial aspects, specific areas in the city turn out to be characterized
by distinctive smells (e.g., parks), while non-central areas
tend to offer bland smellscapes.
• We study the relationship between smell and emotions
(Section 5). Streets with smells of nature and food are associated with positive emotion words (picture tags), while
those with smells for metro and waste are associated with
negative ones.
• We conclude this work by studying the relationship between smell and color (Section 6). We do so by testing
whether a specific color is predominantly present in pictures associated with a given smell. We find that certain
smells are associated with definitive colors (e.g., traffic is
mainly black), while others span a variety of colors (e.g.,
food takes multiple colors).
Introduction
The intensity of modern city life calls for urban places that
offer relief (e.g., greenery), and one way of designing such
places (designing what architects call “restorative environmental experiences”) is to map sensory perceptions at city
scale.
Traditional research on the sensory urban experience has
primarily focused on the visual dimension. Back in the 60s,
Kevin Lynch, for example, showed how what we are able to
see and remember of a city contribute to our ability to navigate it and, ultimately, to our well-being as residents (Lynch
1960). On the other hand, previous research on olfactory and
sonic perceptions has mainly explored negative characteristics (Henshaw 2013; Schafer 1993): urban sounds have been
equated to noise, and smells to nuisance odors, so much so
that both end up being “guilty until proven innocent” (Fox
2006).
To see why, take smell. A comprehensive study of it is
made difficult by the simple fact that odors are hard to capture (Section 2). To partly fix that, we have recently proposed a methodology with which social media data can be
used to capture the smellscape of entire cities (Quercia et al.
2015). After creating a dictionary of smell-related words,
we matched those words with social media metadata (e.g.,
2
Related work
Since we are interested in capturing smell words from picture tags on social media, we look at what computer scientists have done in the area of text mining on social media.
Use of language across time. Dodds et al. analyzed text
on tweets to remotely sense societal-scale levels of happiness and studied how word use changed as a function of
time (Dodds et al. 2011). They collected tweets posted by
over 63 million users in 33 months, and built a tunable hedonometer that analyzed word usage in real-time to gauge
“happiness” levels. They found fascinating temporal variations in happiness over timescales ranging from hours to
years. At year level, they found that, after an upward trend
starting from January to April 2009, average happiness gradually decreased. At month level, they saw average happiness
gradually increased towards the end of each year. At week
level, they found peaks over the weekends, and nadirs on
Mondays and Tuesdays. By looking at individual days, they
c 2016, Association for the Advancement of Artificial
Copyright Intelligence (www.aaai.org). All rights reserved.
309
fected by protracted violence in the context of the “Mexican
Drug War”. The researchers found that, while violence was
on the rise offline, negative emotional expression online was
declining and emotional arousal and dominance were raising: both aspects are known psychological correlates of population desensitization. This suggests that chronic exposure
to violence is indeed associated with signs of desensitization
in social media postings.
were able to show that negative days were seen during unexpected societal trauma such as the 2008 Bailout of the US
financial system and the February 2010 Chilean earthquake.
A year later, Lansdall-Welfare et al. analyzed 484 million
tweets generated by more than 9.8 million users from the
United Kingdom over 31 months. They did so to study
the impact of the economic downturn and social tensions
on the use of language on Twitter (Lansdall-Welfare, Lampos, and Cristianini 2012). More specifically, they studied
the use of emotion words classified in four categories taken
from the tool ‘WordNet Affect’: anger, fear, joy and sadness. This resulted into 146 anger words, 92 fear words,
224 joy words and 115 sadness words. The authors found
that periodic events such as Christmas, Valentine and Halloween were associated with very similar use of words, year
after year. On the other hand, they observed two main negative change-points, one occurring in October 2010, when
the government announced cuts to public spending; and the
other in Summer 2011, when riots broke out in various UK
cities. Interestingly, the increase use of negative emotion
words preceded, not followed, those events, suggesting predictive ability. Golder and Macy studied the 500 million English tweets that 2.4 million users produced during almost
2 years. Based on their hour-by-hour analysis, they found
that offline patterns of mood variations also hold on Twitter:
mood variations were associated with seasonal changes in
day length. People also changed their mood as the working
day progressed and were happier during weekends (Golder
and Macy 2011).
The focus of our work is on urban smell, so summarizing
what computer scientists and urban planners have already
done in the area is in order.
Smell in Computer Science. In 2004, Kaye showed that
the vast majority of work in Computer Science and, more
specifically, in Human-Computer Interaction “involves our
senses of sight and hearing, with occasional forays into
touch.” (Kaye 2004) Since then, work in HCI has focused
on smell technologies that are able to capture and generate
smells. Bodnar and Corbett proposed smell-based notifications and showed they were less disruptive than visual notifications or auditory ones (Bodnar and Corbett 2004). A couple of years later, Brewster et al. designed Olfoto, a photo
tagging tool in which smell was used to search the collection (Brewster, Mcgookin, and Miller 2006). More recently,
to go beyond capturing smell, researchers have explored the
idea of transmitting it as well, and they did so over the Internet (Ranasinghe et al. 2011). More generally, experiences
with smell have profound and multi-faceted implications for
technology (Obrist, Tuch, and Hornbaek 2014).
Use of language across space. Schwartz et al. tested
whether the language used in tweets is predictive of the subjective well-being of people living in US counties (Schwartz
et al. 2013). They did so by collecting a random sample
of tweets in 1,293 US counties. By correlating the word
use with subjective well-being as measured by representative surveys, they found positive correlations with pro-social
activities, exercise, and engagement with personal and work
life, and negative correlations with words of disengagement.
Those findings were in line with existing well-being studies
in the social sciences. Frank et al. went on characterizing the mobility patterns of 180,000 individuals (Frank et al.
2013). In so doing, the researchers were able to characterize
changes in the use of language in relation to people’s mobility. They found that tweets written close to a user’s center of
mass (typical location) are slightly happier than those written 1 km away, which is the distance representative of a short
daily commute to work. Beyond this least happy distance,
they found that the use of positive emotion words increased
logarithmically with distance. This pattern did hold when
they moved to study a user’s radius of movement. The larger
a user’s radius, the more happier words the user tended to
use. Finally, De Choudhury et al. studied whether people
exposed to chronic violence lowered affective responses in
their Twitter posts (De Choudhury, Monroy-Hernández, and
Mark 2014). To this end, they collected all of the Spanish tweets that were mentioning one of the four Mexican
cities of Monterrey, Reynosa, Saltillo, and Veracruz. Between Aug 2010 and Dec 2012, these four cities were af-
Smell in Urban Planning. People are able to detect up to
1 trillion smells (Bushdid et al. 2014), yet city planning focuses on a few bad smells only. City officials entirely rely
on complaints to capture smell and ultimately inform urban
planning. In the research world, smell has been recorded
in a variety of ways (Henshaw 2013): with ‘nose trumpets’
that capture four main olfactory aspects (i.e., odor character,
odor intensity, and duration); with web maps that elicit smell
words from users in a crowdsourcing way; and with sensory
walks (Diaconu 2011) in which groups of people are asked
to walk around the city and record what they smell. Unfortunately, those ways of collecting smell information require
substantial public engagement to be effective.
Missed opportunities. Based on this literature review, one
can see that, in both Computer Science and Urban Planning,
there is no effective solution to capture smell (both its positive and negative aspects) at scale, and such an inability has
likely limited the scholarly production in those two fields.
3
Urban Smell from Social Media
To partly fix that, we have recently proposed a new way
of collecting odor information at scale without requiring
a massive public engagement. This way simply analyzes
data implicitly generated by social media users (e.g., photo
tags) (Quercia et al. 2015), and it unfolds in three main steps:
Step 1: Collecting smell-related words. We conducted
310
has many subcategories (not all shown in the mid-layer of
Figure 1 for space limitation). One of those subcategories
is ‘Fuel’, which, in turn, is associated with some of the 258
smell-related word (with, e.g., ‘gasoline’, ‘petrol’). Interestingly, this categorization (which is purely data-driven)
strikingly resembles previous hand-made classifications in
olfactory research (Henshaw 2013) and has been found
to have ecological validity (Quercia et al. 2015) (i.e., the
main 10 categories tend to be geographically orthogonal
to each other). The dictionary is made available1 for noncommercial purposes to anyone who wishes to translate it in
a language other than the ones already at disposal. The idea
is that, in the long term, dictionaries in different languages
will be freely available to many stakeholders, from artists
and designers to scientists.
Social media data is biased, and that might make it difficult to use that data for urban olfactory analysis. To validate
such a use, we collect data about the three air pollutants –
Nitrogen oxides (NO2), coarse particles (PM10), and fine
particles (PM2.5) – for both London (from King’s College
London API2 ) and Barcelona (from the authors of (Beevers
et al. 2013)). We map and use the pollution data onto 34K
street segments in Barcelona and about 263K in London. A
segment is a street’s portion between two road intersections.
Most segments have a few tags, while a few have many tags
(Figure 3). We then compute the fraction of each segment’s
tags that belong to a given smell category, so each segment
comes with ten smell descriptors (as we have 10 smell categories). We compute and map the fraction of tags in smell
category S at each location l (Figure 2):
#tags in smell category S at location l
fS @l =
(1)
#tags in any smell category at location l
We map neither volume nor density because the use of the
fraction yields the strongest correlation between air pollutants and smell categories (Figure 4).
Then, for further validation, we also collect data about
presence of natural elements and eating places from the
Open Street Map (OSM) database3 . These descriptors identify urban elements that are likely to be associated with certain smells. For example, the OSM natural venue marker is
used to identify natural land features; the vegetation and surface venues include natural elements such as tree, wood, and
grassland; and the cuisine marker is associated with venues
that serve food of any kind, mostly restaurants and markets.
Having those OSM venues at hand, we simply count them
and correlate the counts with our 10-category description
of smell. As one would expect, the presence of nature and
emission smells correlates with OSM natural venues (positively and negatively, respectively), while food smells correlate with OSM cuisine venues (Figure 5).
Figure 1: Urban smell taxonomy. Top-level categories are
in the inner circle; second-level categories, if available, are
in the outer ring; and examples of words are in the outermost ring. For space limitation, in the wheel, only the first
categories (those in the inner circle) are complete, while subcategories and words represent just a sample.
“smellwalks” around seven cities in UK, Europe, and USA.
Locals were asked to walk around their city, identify distinct
odors, and take notes (to, e.g., report smell descriptors). As
a result of those sensory walks, smell-related words were
recorded and classified, resulting in the first urban smell dictionary. The smell dictionary contains words describing the
smell itself (e.g., grassy), which often coincide with the object that emits the smell (e.g., musk, chocolate, pine).
Step 2: Matching smell words with social media. For the
cities of Barcelona and London, we collected geo-referenced
tags from 17M Flickr public photos and 436K Instagram
photos, and 1.7M geo-referenced tweets from Twitter (those
tweets include neither retweets nor direct replies). We then
matched those tags and tweets with the words in the smell
dictionary. In this work, we were able to use the same Flickr
dataset, which is the biggest one and has the highest temporal coverage, spanning a time frame of more than 10 years
(January 2005 to October 2015).
Step 3: Organizing smell words into a dictionary. To
structure this large and apparently unrelated set of words, we
built a co-occurrence graph where nodes are smell words,
and undirected edges are weighted by the number of times
the two words co-occur in the same image. Upon this graph,
we found that ten “clusters of words” (shown in the inner
circle of Figure 1) best describe the graph structure. One
of those clusters is ‘Emissions’, for example. This cluster
4
The when and where of smell
To increase tourism, cities have been encouraged to remove
negatively perceived odors and introduce more pleasant varieties (Dann and Jacobsen 2002). To spot those varieties,
1
http://goodcitylife.org/smellymaps/
http://api.erg.kcl.ac.uk
3
http://wiki.openstreetmap.org/wiki/Category:Keys
2
311
(a) London
(b) Barcelona
10
5
10
4
10
10
10
Spearman correlation
Number of segments
Figure 2: Urban Smell Maps for London and Barcelona. They map the fraction fS @l of smell category S at each street segment
l. Both cities have rich smellscapes in which odors are distributed in predictable ways. Emissions are associated with trafficked
roads, nature with greenery spots, food with central parts of the cities, and animal odors with the zoos.
3
London
Barcelona
2
OSM natural tags
0.3
0.2
0.2
0.1
0.1
Emissions
Nature
0.0
−0.1
−0.2
0
OSM food tags
0.3
Food
0.0
−0.1
20
40
60
80
100
−0.2
0
20
40
60
80
100
Number of tags per segment
1
Figure 5: Correlation between a street segment l’s smell category S ( fS @l ) and the number of OSM venues as the number of picture tags per street segment increases.
0
10 0
10
10
1
10
2
10
3
10
4
Number of tags
Spearman correlation (NO2 )
Figure 3: Number of tags per street segment in London and
Barcelona. Many streets have a few tags, and only a few
streets have a massive number of them. London has 263K
segments with at least one tag, Barcelona 34K.
Emissions
0.8
Nature
0.1
0.7
0.0
0.6
0.5
Figure 6: Temporal autocorrelation RS with a 12-months
time lag (seasonality) of smell category S . Smells with autocorrelation close to 0 are unpredictable (non seasonal),
while those close to 1 (e.g., plant smells) are predictable
(seasonal).
−0.1
0.4
−0.2
0.3
Density
Volume
Norm. Volume
0.2
0.1
0.0
20
40
60
80
100
−0.3
−0.4
20
40
60
80
100
Number of tags per segment
Figure 4: Spearman correlation between the presence of
smell categories fS @l and pollution levels for different ways
of aggregating smell tags: density (number of tags over a
street segment’s area), raw volume count, and relative volume (or fraction).
smells is to determine the smells that show predictable temporal patterns during an entire year. To this end, we consider the time series of the relative frequency of a smell in
each month, and do so for 10 years. These time series represent stationary processes that exhibit different degrees of
seasonality. A standard method to measure the periodicity
of this type of series is the autocorrelation R. This is the
cross-correlation of the temporal signal with itself considering a fixed time window τ. Specifically, we compute the
one needs to know when and where to smell them. We turn
to the question of when first.
Smell of the season. One way of determining seasonal
312
Month
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
Smell
Soil
Soil
Soil
Plants
Plants
Plants
Plants
Traffic
Food
Traffic
Soil
Soil
London
Where
St. James’ Park
St. James’ Park
St. James’ Park
St. James’ park
Ranelagh Gardens
Regent’s Park
Regent’s Park
Piccadilly Circus
Central London
Regent’s Street
Russel Square
Trafalgar Square
Smell
Soil
Soil
Plants
Food
Plants
Traffic
Plants
Food
Food
Food
Food
Food
Barcelona
Where
Parc Guell
Parc Guell
Parc Ciutadella
Boqueria
Parc Guell
Pl. de Espanya
Parc Ciutadella
Boqueria
Boqueria
Boqueria
Boqueria
Boqueria
The results are reported in Table 1: the most frequent categories include trees & soil from November to March, and
flowers & plants from April to July. August and September deviate from the pattern and are characterized by traffic
(August) and food (September). Plants can be smelled in
Ranelagh Gardens (during the Flower Festival in May) and
Regent’s Park. However, not all months are equal: there
might be months that are olfactory more distinctive than
others. To identify them, we compute the Shannon entropy
from the vector of the smell frequencies < fS ,t >S for each
month t. We find that the least distinctive month is January,
while the most distinctive ones are March, April, and May.
Finally, there might be months that are more olfactory pleasant than others. To determine what is pleasant and what is
not, we resort to the literature (Henshaw 2013), which lists
the smells people tend to like and those they tend to dislike
(Table 2). To then derive a pleasantness score out of pictures
for each month t, we compute the pleasure score, which is
the z-score of the fraction of pleasant tags minus the z-score
for the unpleasant ones:
Table 1: The smell of the month, and where to experience it
best.
Pleasant smells
bread, baked, baked goods, coffee, coffees, aftershave, cut
grass, grass, grassy, floral, flower, flowers, flowershop, flowery,
lavender, lilies, lily, magnolia, rose, rosey, tulip, tulips, violet,
violets, baby, babies, child, children, sea, seaside, countryside,
cedar, cedarwood, conifer, dry grass, earth, earthy, eucalyptus,
ground, leafy, leaves, old wood, pine, sandalwood, soil, tree,
trees, wood, woodlands, woody, petrol, diesel, fuel, gasoline,
soap powder, soap
f pleasant,t − μ pleasant
funpleasant,t − μunpleasant
−
σ pleasant
σunpleasant
(4)
Where μ pleasant|unpleasant and σ pleasant|unpleasant are the average and standard deviation of the fraction of tags reflecting
pleasant (unpleasant) smells across all locations and all calendar months (considering a 10-years statistics). The higher
the pleasantness score, the higher the concentration of pleasant smells over unpleasant ones. Pleasantness is zero when
the number of pleasant tags and that of unpleasant ones are
both equal to their average values. We plot the pleasantness
z-scores for twelve months (Figure 8), and find that the peak
is registered in May, which turns out to be the best smelling
month of the year.
z pleasure,t =
Unpleasant smells
flatulence, fart, vomit, dog shit, dogshit, excrement, faeces,
fart, farts, feces, manure, shit, cigarette smoke, cigarette,
cigarettes, cigar, cigars, smoker, tabacco, tobacco, pee, piss,
ammonia, urine, public toilet, public toilets, toilet, toilets,
urinal, urinals, gone-off milk, fish, rotten fish, rotten food,
rotten, rotten fruit, rotten fruits, putrid, bus, buses, car, cars,
exhaust, traffic, fume, fumes, body odour, body odor, sweat,
sweaty, dirty clothes
Table 2: List of smells people generally find pleasant and
those they find unpleasant.
autocorrelation as follows:
E ( fS ,t − μS ) · ( fS ,t+τ − μS )
RS ,τ =
σ2S
The where of smell. Knowing the smells of the month, we
now turn to determine where to perceive them. We find that,
in London, soil smells are best experienced near Trafalgar
Square, Russel Square, and St James’ Park (Table 1). We
find similar results for Barcelona, with a greater emphasis
for food smells though. These are best experienced in La
Boqueria, which is the main food market in the center of the
city. We then identify the most (un)pleasant locations (street
segments) by computing z pleasure@l for each location l:
(2)
where E is the expected value, fS ,t is the fraction of the smell
category S at month t, τ is the time window we set at 12
months (as we are interested in yearly seasonality), and μS
and σ2S are the average and variance of the whole time series
for the smell category S . R ranges in [−1, 1] (Figure 6). We
find that the most predictable (seasonal) smells are those of
plants (R = 0.64) followed by soil (R = 0.58) and food (R =
0.32). All other smells are non-seasonal, the one closest to
0 being the smell of waste (R = −0.02).
f pleasant@l − μ pleasant
funpleasant@l − μunpleasant
−
σ pleasant
σunpleasant
(5)
We then map those values (Figure 7) and see that parks tend
to have the most pleasurable smells, while main roads are
infested with the least pleasurable.
z pleasure@l =
Smell of the month. Some smells characterize not only an
entire season but also specific months. To show that, for each
month t, we compute the most frequent smell categories, that
is, we rank each smell category S by fS ,t and choose the one
at the top:
fS ,t =
#tags in smell category S at month t
#tags in any smell category at month t
5
The emotion of smell
Looking at a location through the lens of social media makes
it possible to study that location from different points of
views. So far we have studied the spatio-temporal dynamics
of smell. Yet, the sense of smell has a highly celebrated link
with other aspects, most notably with emotions. The nose
(3)
313
(a) London
(b) Barcelona
Figure 7: Maps of the olfactory pleasantness of street segments in London and Barcelona. They report the z-scores of pleasant
smells z pleasure@l at each street segment l. The color of a street goes from green (very pleasant) to red (very unpleasant).
Pleasantness scores can be further explored at http://goodcitylife.org/smellymaps by selecting a street segment of interest.
Figure 8: Olfactory pleasantness z pleasure,t of each month
based on 10-year statistics.
Figure 9: Pearson correlation between presence of smell
category fS @l and presence of positive emotion words
z sentiment@l . Values are computed considering street segments
with at least 30 smell tags. The p-value is always < 0.01.
z sentiment@l =
Spearman(LIWC,smell)
has direct access to the amygdala, the part of the brain that
controls emotional response (Gilbert 2008). As such, smells
have a considerable effect on our feelings and our behavior.
Therefore, we set out to study the relationship between
the smellscape and emotions on our data. To do so, we need
to have a lexicon of emotion words. We use two of them:
the “Linguistic Inquiry Word Count” (LIWC) (Pennebaker
2013), that classifies words into positive and negative emotions, and the “EmoLex” word-emotion lexicon (Mohammad and Turney 2013), that classifies words into eight primary emotions based on Plutchik’s psycho-evolutionary theory (Plutchik 1991) (i.e., anger, fear, anticipation, trust, surprise, sadness, joy, and disgust).
From those two lexicons, we use the polarity of positive
and negative words to compute the sentiment score by subtracting the z-score of the negative tags from the z-score of
the positive ones:
0.4
0.2
Food
Metro
0.0
Nature
Waste
−0.2
−0.4
0
20
40
60
80
100
Number of tags per segment
Figure 10: Spearman correlation between presence of smell
category fS @l and sentiment score z sentiment@l as the number
of tags per street segment increases. The four smell categories with the strongest correlations are shown. The pvalue (across all street segments) is always < 0.01.
f positive@l − μ positive fnegative@l − μnegative
−
(6)
σ positive
σnegative
vidual emotions:
#tags emotion category E at location l
fE@l =
#tags in any emotion category at location l
From EmoLex, we compute the z-score of its eight indi-
314
(7)
Spearman(LIWC,smell sentiment
6
0.6
We have seen that there is an association between emotional
and olfactory layers. Emotions are triggered not only by the
sense of smell but also by the sense of sight. What we see
impacts how we feel. When detecting images, our retina
generates nerve impulses for varying colors. In its most
basic form, our vision revolves around colors and, consequently, early studies have related colors to emotions.
In marketing, colors are widely used to influence consumers’ emotions and perceptions. Despite cross-cultural
differences, there are significant cross-cultural similarities
regarding which emotional states people associate with different colors. For example, the color red is often perceived
as strong and active (Widermann, Barton, and Hill 2011),
tones of black lead to feelings of grief and fear, while green
tones are often associated with good taste (Aslam 2006).
We therefore explore this last relationship: that between
smellscape and colors. As a first step, we need to extract
colors from our pictures4 (Figure 13). We do so not from
the images themselves (to avoid the noise introduced by the
images with multiple colors) but from tags. To textually extract colors from tags, we build a color term dictionary by
grouping 249 color nuances into ten main colors (collating
the colors into fewer chromatic categories greatly reduces
spurious matches)5 : black, blue, yellow, gray, green, orange,
red, violet, white, and yellow.
Technically, having color c at hand, we compute the
strength of its association with smell s based on color-smell
co-occurrences, normalized by the total for that color:
0.5
0.4
0.3
0.2
0.1
0.0
0
50
100
150
The color of smell
200
Number of tags per segment
Figure 11: Spearman correlation between sentiment score
z sentiment@l and presence of pleasant smell (tags) z pleasure@l
as the number of tags per street segment (x-axis) increases.
The p-value is always < 0.01
Emotions and smells. After computing the Pearson correlations between the presence of each smell category S and
LIWC sentiment scores (Figure 9), we find that positive sentiment tags are found in streets with food and nature smells,
while negative sentiment tags are found in streets with waste
and metro smells. To be significant, those correlations do not
require a large number of tags per street: it depends on the
smell categories but 50 tags are usually enough (Figure 10).
The same results are obtained when the sentiment score is
computed from the classification of EmoLex. Furthermore,
this latter lexicon allows for studying finer-grained emotions
as it considers eight emotional constructs. We correlate the
fraction fS @l of tags in smell category S at each location and
the fraction of tags in each emotion category E and show
the results in Figure 12. We observe that waste correlates
positively with disgust and sadness but negatively with joy.
Similarly, where emission-related smells are present, joyrelated words are absent. Interestingly, emission smells positively correlate with trust and fear (whose combination is
interpreted as submission by Plutchik’s theory). Those correlations are especially strong in dense urban areas in which
the fear generated by intense traffic likely mixes with the
trust and safety generated by the presence of crowds (Speck
2012).
strength s,c = pcs
pc +p s
pcs
c pc +p s
(8)
where pcs is the number of photos in which c and s cooccur, pc is the number of photos associated with color c,
and p s is the number of photos associated with smell s. To
make strength scores comparable across smell categories,
cs
we divide the ratio by the rescaling factor c pcp+p
. In coms
puting the strength score, we make sure to consider only the
(smell,color) associations for which we have at least 10 pictures (i.e., pcs ≥ 10), effectively avoiding spurious associations.
Instead of determining the color of each individual smell
(which would suffer from data sparsity), it would be more
informative to determine the color of an entire smell category. To do so, we aggregate the fine-grained color-smell
associations. A smell category consists of individual smell
words. For each smell word s (e.g., violet), we determine
the most representative color by selecting color c (c strongest )
with the highest strength s,c (e.g, purple). To then determine
the color of the smell category S (e.g., nature), we sum
the strengths for all colors in that category: strengthS ,c =
∀s∈S strength s,c strongest . We do so by considering only the
(smell,color) associations for which we have at least 10 pictures. The color associations (Figure 14) meet expectations:
traffic is associated with black and red (likely from road,
Emotions and pleasant smells. In Section 4, to find the
most olfactory pleasant month, we have been able to distinguish pleasant from unpleasant smells based on the literature
(Table 2). One could reasonably hypothesize that areas with
(un)pleasant smells are characterized by specific emotions.
To verify that, we compute the Pearson correlation between
a street segment’s (un)pleasant smells (as per Formula 4 in
Section 4) and the segment’s sentiment. We find that, indeed, locations with pleasant smells tend to be associated
with positive emotion tags (with correlation r up to 0.50),
while locations with unpleasant smells tend to be associated
with negative ones. The correlations change depending on
the number of tags per street segment but are quite stable
after 150 tags (Figure 11).
4
5
315
We discard black and white photos.
www.farb-tabelle.de/en/table-of-color.htm
Figure 12: Correlation between the fraction fS @l of tags in each smell category S and the fraction fE@l of tags in each Plutchik’s
emotion category E. Positive correlations are in green, and negative ones are in red. All correlations are statistically significant
at the level of p < 0.01.
Figure 13: Examples of pictures in which a smell tag and
a color tag co-occur. The first row has (from left to right)
tree+violet, food+brown, and traffic+gray. The second row
has tree+green/brown, food+brown, and traffic+red.
Figure 14: Bipartite graph of smell-color associations (best
seen in color). The ticker a line, the higher the association strength strengthS ,c between smell category S and color
c. The percentage next to each color is the average fraction of that color across all smells. For example, black
characterizes any smell 29% of the times. Only the five
smells with the lowest entropy of color mixture are shown.
An interactive version is available at http://goodcitylife.org/
smellymaps/chromosmell.
smoke, traffic lights and red buses); industrial smells with
black as well; trees and soil with green; flowers with green,
violet, and orange; and food with brown and orange. To allow for reproducibility and the re-use of our results, we spell
out all the associations strengths in Figure 15.
We are not the first to relate colors to smell. It has been
repeatedly shown that there is a reliable multi-sensory interaction between odors and colors (e.g., between the color yellow and the odor of bergamot). By asking study participants
to associate those two things, researchers have found consistent associations (Gilbert, Martin, and Kemp 1996). To go
beyond explicit matches, Dematte et al. used an indirect association measure (called implicit association test (Anthony
G. Greenwald and Debbie E. McGhee and Schwartz JL.
1998)) and showed that color-odor associations are both systematic and robust (Dematte, Sanabria, and Spence 2006).
All those studies have focused on associations made by individuals; we have studied the same associations at geographic level. We have done so because our insights can
inform the design of smell-related technologies. As a shortterm example, consider mapping. When showing a specific
smell on the city map or on any user interface, it is impor-
316
smellscapes both spatially and temporally. Olfactory researchers have so far focused on negative characteristics of
smell. Now, based on unobtrusive data capturing from social media users, the same researchers could move forward
and study the positive role that odors can play in the environmental urban experience. We have also shown for the first
time how sensorial perceptions can be mapped to orthogonal
dimensions like the ones of emotions and colors.
All these aspects are not incremental. On the contrary,
they pertain to totally different domains than that of smell.
Think about a park. As for smell, it falls into the nature category. As for emotions, instead, it is multi-faceted: different
parts of the park speak to opposite emotions (e.g., lavender is calming, and fresh grass is energizing). Then, as for
time, the park drastically changes across seasons. Finally, as
for colors, the chromatic reality implicitly suggested by the
pictures goes beyond the obvious description of the stereotypical green park.
Figure 15: Matrix of smell-color associations. The darker,
the stronger the association. Each number is strengthS ,c ,
which reflects the association between color c and smell category S . Smells are sorted in ascending order of entropy
of color mixture (entropy has nothing to do with association strength and is the Shannon entropy of each row in the
matrix). Smells at top are associated with very few colors
(e.g., trees are predominantly associated with green), while
smells at the bottom are associated with a variety of colors
(e.g., flowers are colorful).
Practical Implications. Our work might help a variety of
stakeholders. Urban planners could go beyond the attempt
of mitigating bad odors and their potential negative impact
and could use, instead, social media tools to monitor the
whole spectrum of emotion associated with the smellscape.
This might inform policies to incentivize the development
of areas with pleasurable smells. On a similar note, the
tourism industry could capitalize on olfactory opportunities by facilitating the discovery and exploration of places
that are not conventionally included in touristic tours. Computer scientists have worked on way-finding tools that suggest not only shortest routes between two points in the city
but also short ones that are beautiful (Quercia, Schifanella,
and Aiello 2014) or visually distinctive (Noulas et al. 2012;
Van Canneyt et al. 2011). Now they could look into
recommending routes that are olfactorily pleasant, as our
work provides a principled and scalable method to capture
smell. Interaction designers could inform their work with
the knowledge of how color, smell, and emotions are interrelated and, as such, know which color to use on, say,
a map when graphically representing smell. As a proof-ofconcept, we built an interactive map6 that allows to navigate
the smellscape of central London. The interface has been
selected as one of the finalist projects in the Insight Competition by CartoDB7 and is exhibited at the Storefront for Art
and Architecture in New York City. As for the last stakeholder, the general public could now nurture a critical voice
for the positive and negative role that smell has to play in the
city.
tant to know which color to use. As a long-term example,
consider virtual reality. A multi-sensory virtual reality experience should effectively match the visual (including colors) and the olfactory experience. In fact, it has been shown
that, to be effective, smell must match context (studies refer
to that as the “congruency” problem (Gilbert 2008)). To see
why, consider the researchers who explored the combined
effect of smell and music in a gift shop (Mattila and Wirtz
2001): consumer satisfaction increased when the store had
(low-arousal) lavender and relaxing music or (high-arousal)
grape-fruit and energizing music; by contrast, no effect was
registered when there was a mismatch between music and
smell.
7
Discussion
Before concluding, we briefly discuss our work’s limitations, theoretical implications, and practical implications.
Limitations. This work has considered the “average” urban smellscape. Yet, smell is mediated by individual factors
(e.g., age, gender), geographic factors (e.g., climatic conditions), and contextual factors (e.g., city layout) (Henshaw
2013). Also, not all odors are the same. It is difficult to capture fleeting odors from social media, not least because those
odors are localized in space and time. Our analysis, instead,
is able to capture what olfactory researchers call base smell
notes (uniformly distributed across the city) and mid-level
notes (localized in specific areas of the city).
8
Conclusion
In Japan, one hundred sites have been declared as protected
because of their ‘good fragrance’. By contrast, in the rest
of the world, environmental smells have received little attention. Our urbanization age, however, results into closer
proximity between people and activities and, as such, faces
new olfactory challenges. To tackle those challenges, we
Theoretical Implications. This work has shown that social media could be used to capture insights about urban
6
7
317
http://goodcitylife.org/smellymaps/
http://blog.cartodb.com/insight-finalists
need to capture the complex olfactory fragments of our
cities. We have used social media to do so and studied
the interplay between dynamics of different nature: spatiotemporal, emotional, and chromatic. To build upon this
work, we are currently exploring three main directions.
First, we are working on a smell app to capture the fleeting
odors that cannot be extracted from social media. Second,
we have teamed up with epidemiologists to determine under
which conditions social media complements the use of extremely costly air sampling devices. Third, we are designing
a study to assess whether it is possible to change the behavior of city dwellers by making urban smell visible. The goal
of this work has been to make visible what is normally not
and, ultimately, to make it possible to create more fulfilling,
humanistic, and sustainable urban environments.
Gilbert, A. N.; Martin, R.; and Kemp, S. E. 1996. Cross-modal
correspondence between vision and olfaction: the color of smells.
American Journal of Psychology 11(3):335–51.
Gilbert, A. 2008. What the Nose Knows: The Science of Scent in
Everyday Life. Crown/Archetype.
Golder, S. A., and Macy, M. W. 2011. Diurnal and Seasonal Mood
Vary with Work, Sleep, and Daylength Across Diverse Cultures.
Science 333(6051):1878–1881.
Henshaw, V. 2013. Urban Smellscapes: Understanding and Designing City Smell Environments. Routledge.
Kaye, J. J. 2004. Making Scents: Aromatic Output for HCI. ACM
Interactions 11(1):48–61.
Lansdall-Welfare, T.; Lampos, V.; and Cristianini, N. 2012. Effects
of the Recession on Public Mood in the UK. In Proceedings of
ACM WWW, 1221–1226.
Lynch, K. 1960. The Image of the City. Publication of the Joint
Center for Urban Studies. Mit Press.
Mattila, A. S., and Wirtz, J. 2001. Congruency of scent and music
as a driver of in-store evaluations and behavior. Journal of Retailing 77(2):273–289.
Mohammad, S. M., and Turney, P. D. 2013. Crowdsourcing
a word–emotion association lexicon. Computational Intelligence
29(3):436–465.
Noulas, A.; Scellato, S.; Lathia, N.; and Mascolo, C. 2012. A random walk around the city: New venue recommendation in locationbased social networks. In Proceedings of ASE/IEEE SocialCom,
144–153. IEEE.
Obrist, M.; Tuch, A. N.; and Hornbaek, K. 2014. Opportunities for
Odor: Experiences with Smell and Implications for Technology. In
Proceedings of ACM CHI, 2843–2852.
Pennebaker, J. 2013. The Secret Life of Pronouns: What Our Words
Say About Us. Bloomsbury.
Plutchik, R. 1991. The emotions. University Press of America.
Quercia, D.; Aiello, L. M.; Schifanella, R.; and McLean, K. 2015.
Smelly Maps: The Digital Life of Urban Smellscapes. In Proceedings of AAAI ICWSM.
Quercia, D.; Schifanella, R.; and Aiello, L. M. 2014. The shortest path to happiness: Recommending beautiful, quiet, and happy
routes in the city. In Proceedings of ACM HT.
Ranasinghe, N.; Karunanayaka, K.; Cheok, A. D.; Fernando, O.
N. N.; Nii, H.; and Gopalakrishnakone, P. 2011. Digital taste and
smell communication. In Proceedings of BodyNets.
Schafer, R. 1993. The Soundscape: Our Sonic Environment and
the Tuning of the World. Inner Traditions/Bear.
Schwartz, H. A.; Eichstaedt, J. C.; Kern, M. L.; Dziurzynski, L.;
Lucas, R. E.; Agrawal, M.; Park, G. J.; Lakshmikanth, S. K.; Jha,
S.; Seligman, M. E. P.; and Ungar, L. H. 2013. Characterizing
Geographic Variation in Well-Being Using Tweets. In Proceedings
of AAAI ICWSM.
Speck, J. 2012. Walkability: How downtown can save america one
step at a time.
Van Canneyt, S.; Schockaert, S.; Van Laere, O.; and Dhoedt, B.
2011. Time-dependent recommendation of tourist attractions using
Flickr. In Proceedings of BNAIC, 255–262.
Widermann, D.; Barton, R. A.; and Hill, R. A. 2011. Evolutionary perspectives on sport and competition. Applied Evolutionary
Psychology 290–307.
Acknowledgments
We thank researchers at the Centre for Research in Environmental
Epidemiology (CREAL) for providing air quality data for the city
of Barcelona.
References
Anthony G. Greenwald and Debbie E. McGhee and Schwartz JL.
1998. Measuring individual differences in implicit cognition: The
implicit association test. Journal of Personality and Social Psychology 74(6):1464–80.
Aslam, M. 2006. Are You Selling the Right Colour? A Crosscultural Review of Colour as a Marketing Cue. Journal of Marketing Communications 1(12):15–30.
Beevers, S. D.; Kitwiroon, N.; Williams, M. L.; Kelly, F. J.; Anderson, H. R.; and Carslaw, D. C. 2013. Air pollution dispersion
models for human exposure predictions in London. Journal of Exposure Science and Environmental Epidemiology 23:647–653.
Bodnar, A., and Corbett, R. 2004. AROMA: Ambient awaReness through Olfaction in a Messaging Application - Does Olfactory Notification Make ’Scents’? In Proceedings of ACM ICMI.
Brewster, S. A.; Mcgookin, D. K.; and Miller, C. A. 2006. Olfoto:
Designing a smell-based interaction. In Proceedings of ACM CHI.
Bushdid, C.; Magnasco, M. O.; Vosshall, L. B.; and Keller, A.
2014. Humans Can Discriminate More than 1 Trillion Olfactory
Stimuli. Science 343(6177):1370–1372.
Dann, G. M., and Jacobsen, J. K. S. 2002. Leading the Tourist by
the Nose. In The Tourist as a Methaphor of the Social World.
De Choudhury, M.; Monroy-Hernández, A.; and Mark, G. 2014.
“Narco” Emotions: Affect and Desensitization in Social Media
During the Mexican Drug War. In Proceedings of ACM CHI, 3563–
3572.
Dematte, M. L.; Sanabria, D.; and Spence, C. 2006. Cross-Modal
Associations Between Odors and Colors. Oxford Journal of Chemical Senses 31(6):531–538.
Diaconu, M. 2011. Senses and the City: An Interdisciplinary Approach to Urban Sensescapes. LIT Verlag.
Dodds, P. S.; Harris, K. D.; Kloumann, I. M.; Bliss, C. A.; and
Danforth, C. M. 2011. Temporal Patterns of Happiness and Information in a Global Social Network: Hedonometrics and Twitter.
PLoS ONE 6(12).
Fox, K. 2006. The Smell Report. Social Issues Research Centre.
Frank, M. R.; Mitchell, L.; Dodds, P. S.; and Danforth, C. M. 2013.
Happiness and the Patterns of Life: A Study of Geolocated Tweets.
Scientific Reports 3(2625).
318